CN115953031A - Risk prediction model training method and device, computer-readable storage medium - Google Patents

Abstract

The present disclosure relates to a training method and device for a risk prediction model, a computer-readable storage medium, and relates to the field of artificial intelligence. The training method of the risk prediction model includes: obtaining a plurality of training samples, wherein each training sample includes a risk label and user characteristics, and the risk label of at least one training sample in the plurality of training samples includes a first risk generated in a first time window label and the second risk label generated in the second time window, the length of the first time window is greater than the length of the second time window; for each training sample, a risk prediction model is used to generate a risk prediction result according to user characteristics; according to multiple The risk prediction results of the training samples and the corresponding risk labels are used to train the risk prediction model. According to the training method of the risk prediction model disclosed in the present disclosure, the accuracy of the risk prediction model in long-term risk prediction and short-term risk prediction can be improved simultaneously.

Description

Risk prediction model training method and device, computer-readable storage medium

technical field

The present disclosure relates to the field of Internet technologies, and in particular to a training method of a risk prediction model, a risk prediction method and device, electronic equipment, and a computer-readable storage medium.

Background technique

With the rapid development of the Internet and communication technology, traditional risk control methods have gradually been unable to support risk prediction needs, while Internet technology's intelligent processing of multi-dimensional and large amounts of data, and batch standardized execution processes are more suitable for risks in the era of information development. Control development requirements. In the field of Internet technology, by establishing various risk prediction models, learning user behavior patterns, and mining the value of massive data, the purpose of reasonably avoiding risks can be achieved.

The risk in the Internet refers to the risk of losses caused by users failing to fulfill the obligations in the contract. In related technologies, by establishing a risk prediction model and learning rules from user characteristics, it is predicted whether a user has a risk of default, thereby performing risk control and risk reminders to users, and avoiding the risk of user default.

Contents of the invention

According to a first aspect of the present disclosure, a method for training a risk prediction model is provided, including:

Obtaining multiple training samples, wherein each training sample includes a risk label and user characteristics, and the risk label of at least one training sample in the multiple training samples includes a first risk label generated in a first time window and a risk label generated in a second time window The second risk label, the length of the first time window is greater than the length of the second time window;

For each training sample, a risk prediction model is used to generate a risk prediction result according to user characteristics, wherein the risk prediction result of the at least one training sample includes the prediction result of the first risk prediction task corresponding to the first risk label and the second risk The prediction result of the second risk prediction task corresponding to the label;

According to the risk prediction results and corresponding risk labels of multiple training samples, the risk prediction model is trained.

In some embodiments, the risk prediction model is trained according to the risk prediction results and corresponding risk labels of multiple training samples, including:

For each training sample in the at least one training sample, the at least one training sample is calculated according to the first risk label, the second risk label, the prediction result of the first risk prediction task, and the prediction result of the second risk prediction task The loss function for each training sample in ;

According to the loss function of multiple training samples, the risk prediction model is trained.

In some embodiments, for each training sample in the at least one training sample, according to the first risk label, the second risk label, the prediction result of the first risk prediction task and the prediction result of the second risk prediction task, calculate The loss function of each training sample in the at least one training sample includes:

For each training sample in the at least one training sample, calculate a first loss function for each training sample in the at least one training sample according to the first risk label and the prediction result of the first risk prediction task;

For each training sample in the at least one training sample, calculate a second loss function for each training sample in the at least one training sample according to the second risk label and the prediction result of the second risk prediction task;

A loss function for each of the at least one training samples is calculated based on the first loss function and the second loss function for each of the at least one training samples.

In some embodiments, according to the first loss function and the second loss function of each training sample in the at least one training sample, calculating the loss function of each training sample in the at least one training sample includes:

The loss function of each training sample in the at least one training sample is calculated according to the sum of the first loss function and the second loss function of each training sample in the at least one training sample.

In some embodiments, the user features of the at least one training sample are extracted at a specified time, the first time window is a time window from the specified time to the first time, and the second time window is the time from the specified time to the second time window.

In some embodiments, the training method of the risk prediction model also includes at least one of the following:

Fill missing values in user features with default values;

Fill missing values in user features with mean filling;

Fill missing values in user features with the nearest complement method;

Fill missing values in user features with the nearest neighbor method;

Fill missing values in user characteristics with multiple imputation.

In some embodiments, the training method of the risk prediction model further includes screening out at least one of the following user characteristics:

User features whose missing rate is greater than the first threshold;

User features whose discrimination degree to the training samples is lower than the second threshold;

User features with variance greater than a third threshold.

In some embodiments, the training method of the risk prediction model also includes:

Discretize user features.

In some embodiments, the training method of the risk prediction model also includes:

The risk prediction model is validated using validation samples, where the validation samples are generated at different times from the training samples.

In some embodiments, the risk prediction model is a multi-task deep neural network model or a tree model.

According to a second aspect of the present disclosure, a risk prediction method is provided, including:

According to the user characteristics of the target user, a risk prediction model is used to generate a risk prediction result for the target user, wherein the risk prediction model is trained according to the risk prediction model training method described in any embodiment of the present disclosure.

According to a third aspect of the present disclosure, a training device for a risk prediction model is provided, including:

The acquisition module acquires a plurality of training samples, wherein each training sample includes a risk label and user characteristics, and the risk label of at least one training sample in the plurality of training samples includes the first risk label generated in the first time window and the first risk label generated in the second time window. The second risk label generated by the time window, the length of the first time window is greater than the length of the second time window;

The generation module is for each training sample, according to user characteristics, using the risk prediction model to generate a risk prediction result, wherein the risk prediction result of the at least one training sample includes the prediction result of the first risk prediction task corresponding to the first risk label and The prediction result of the second risk prediction task corresponding to the second risk label;

The training module trains the risk prediction model according to the risk prediction results of multiple training samples and the corresponding risk labels.

According to a fourth aspect of the present disclosure, there is provided a risk prediction device, including: a generating module configured to use a risk prediction model to generate a risk prediction result for a target user according to the user characteristics of the target user, wherein the risk prediction model is based on The training device of the risk prediction model described in any embodiment of the present disclosure is trained.

According to a fifth aspect of the present disclosure, there is provided an electronic device, comprising:

storage; and

A processor coupled to the memory, the processor configured to execute the method for training a risk prediction model according to any embodiment of the present disclosure, or according to any one of the present disclosure, based on the instructions stored in the memory. The risk prediction method described in the embodiment.

According to a sixth aspect of the present disclosure, there is provided a computer-readable storage medium, on which computer program instructions are stored. When the instructions are executed by a processor, the training of the risk prediction model according to any embodiment of the present disclosure is implemented. method, or the risk prediction method according to any embodiment of the present disclosure.

Description of drawings

The accompanying drawings, which constitute a part of this specification, illustrate the embodiments of the disclosure and together with the description serve to explain the principles of the disclosure.

The present disclosure can be more clearly understood from the following detailed description with reference to the accompanying drawings, in which:

FIG. 1 shows a flowchart of a method for training a risk prediction model according to some embodiments of the present disclosure;

FIG. 2 shows a construction method of risk prediction labels and user characteristics according to some embodiments of the present disclosure;

FIG. 3 illustrates a data preprocessing method according to some embodiments of the present disclosure;

FIG. 4 shows a flowchart of training a risk prediction model using risk prediction results according to some embodiments of the present disclosure;

FIG. 5 illustrates a method of calculating a loss function according to some embodiments of the present disclosure;

6 shows a block diagram of a training device for a risk prediction model according to some embodiments of the present disclosure;

Fig. 7 shows a block diagram of an electronic device according to other embodiments of the present disclosure;

Figure 8 shows a block diagram of a computer system for implementing some embodiments of the present disclosure.

Detailed ways

Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that relative arrangements of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present disclosure unless specifically stated otherwise.

At the same time, it should be understood that, for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

The following description of at least one exemplary embodiment is merely illustrative in nature and in no way intended as any limitation of the disclosure, its application or uses.

Techniques, methods and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, such techniques, methods and devices should be considered part of the description.

In all examples shown and discussed herein, any specific values should be construed as exemplary only, and not as limitations. Therefore, other examples of the exemplary embodiment may have different values.

It should be noted that like numerals and letters denote like items in the following figures, therefore, once an item is defined in one figure, it does not require further discussion in subsequent figures.

In related technologies, when establishing a risk control model, each user is used as a sample, and the time when the user applies is used as the observation point, and a period of time before the observation point is used as the observation period, which is used to extract the user's behavior during the observation period Data, construct the feature X of the sample.

The time window from the observation point to the performance point is called the performance period. The performance period is the time period for monitoring the user's performance at the observation point. According to the user's performance in this time window, the risk label Y is classified into "0" and "1". "0" and "1" indicate whether the user will default, lose contact, etc. in the future. Compared with the observation point, the performance period is the future time, and the acquisition of risk labels needs to wait for more than several months, and the number of samples with a long performance period is less. Therefore, there is a certain hysteresis in the label Y, and there is a phenomenon of delayed feedback.

Using tags with a longer performance period to model, the user's behavior can be fully represented, and the user's default risk is more fully exposed, but the number of samples that can be used is less, which affects the training effect and accuracy of the model. Modeling with risk labels with a shorter performance period, although more samples can be used, but labels with a short performance period cannot fully expose the user's long-term default risk, so the trained model cannot learn the user's behavior pattern well , the prediction of the possibility of the user's long-term breach of contract is not accurate enough, resulting in losses.

When selecting samples to train the model, it is difficult to maintain a balance between the integrity of the samples and the quality of the samples.

In addition, the model trained with a certain label can only predict the user's performance on the risk corresponding to the label, but it does not have generalization.

The disclosure provides a risk prediction model training method, risk prediction method and device, electronic equipment, and a computer-readable storage medium, which can simultaneously improve the accuracy of the risk prediction model in long-term risk prediction and short-term risk prediction.

Fig. 1 shows a flowchart of a method for training a risk prediction model according to some embodiments of the present disclosure.

As shown in Fig. 1, the training method of the risk prediction model includes step S1-step S3. In some embodiments, the method for training a risk prediction model is executed by a training device for a risk prediction model.

In step S1, a plurality of training samples are obtained, wherein each training sample includes a risk label and user characteristics, and the risk label of at least one training sample in the plurality of training samples includes the first risk label generated in the first time window and the first risk label generated in the first time window. For the second risk label generated by the second time window, the length of the first time window is greater than the length of the second time window.

FIG. 2 illustrates a construction method of risk prediction tags and user characteristics according to some embodiments of the present disclosure.

As shown in Figure 2, the user features of at least one training sample are extracted at a specified time, the first time window is the time window from the specified time to the first time, and the second time window is the time window from the specified time to the second time.

A period of time before the specified time (that is, the observation point) is an observation period, located on the left side of the time axis, and is a time interval for observing user characteristics. For example, at a specified time, user characteristics generated during the observation period are extracted.

Those skilled in the art should know that the specified time may be a time interval, that is, the time when users submit applications, and users who submit applications within this time interval constitute samples for modeling.

A window of time after the observation point is used to generate risk labels. According to some embodiments of the present disclosure, among the samples used for training the model, part of the training samples simultaneously have labels generated by multiple time windows of different lengths. For example, the training sample A has both the first risk label Y ₁ and the second risk label Y ₂ . Y ₁ and Y ₂ are respectively used to indicate whether different types of the first risk and the second risk occur, that is, the first risk label and the second risk label correspond to different risk prediction tasks respectively. For example, the first risk is that the user defaults for 90 days, and the second risk is that the user defaults for 30 days. Y ₁ takes values 1 and 0 to indicate whether the first risk occurs, and Y ₂ takes values 1 and 0 to indicate whether the second risk occurs. The first risk label Y ₁ corresponds to the first time window (performance period), that is, the time interval from the specified time to the first time, and its length is T ₁ . The second time window (expression period) corresponding to the second risk label Y ₂ , that is, the time interval from the specified time to the second time, has a length of T ₂ , where T ₁ >T ₂ .

A training sample can have more than two risk labels at the same time, for example, Y ₁ , Y ₂ , Y ₃ , . . . Y _N and other multiple labels, which is not limited in this disclosure.

In addition, some training samples can have only one risk label. For example, the set of samples (training set) used to train the model also includes training samples B and D, training sample B has a risk label Y ₃ , training sample C has a risk label Y ₄ , and the length of the time window corresponding to Y ₃ is greater than Y ₄ The time window Y ₄ . That is, the set of training samples includes training samples with risk labels of different lengths.

In step S2, the risk prediction model is used to generate a risk prediction result for each training sample according to user characteristics, wherein the risk prediction result of at least one training sample includes the prediction result of the first risk prediction task corresponding to the first risk label The prediction result of the second risk prediction task corresponding to the second risk label.

For example, input all samples into the model to obtain the prediction results corresponding to each risk label of the samples.

If a sample has multiple labels, multiple prediction results of multiple tasks corresponding to multiple labels are generated, and the risk prediction results of at least one training sample include the prediction results of the first risk prediction task corresponding to the first risk label and the second The prediction result of the second risk prediction task corresponding to the risk label. The training sample A has both the first risk label Y ₁ and the second risk label Y ₂ , and the corresponding prediction result p ₁ and the second risk label p ₂ of the first risk prediction task are generated.

The first risk label and the second risk label respectively correspond to different risk prediction tasks. The risk prediction model can simultaneously learn the first risk prediction task corresponding to the first risk label and the second risk prediction task corresponding to the second risk label, and obtain an overall optimal solution on multiple risk prediction objectives. Compared to training a single-task model using training samples with only one label, using multi-task learning techniques in risk prediction models has the following advantages:

(1) Since multiple risk prediction tasks corresponding to multiple risk labels (generated by time windows of different lengths) have certain correlations and have their own noises, multi-task learning can achieve implicit data enhancement;

(2) Multi-task learning can help the risk prediction model to focus on important user characteristics and risk prediction tasks. Other risk prediction tasks provide additional evidence for the relevance or irrelevance of these important user characteristics, and the enhanced model is aimed at The learning effect of each risk prediction task;

(3) It can make the risk prediction model learn a generalized user feature representation, which plays the role of implicit regularization.

Before training the model, preprocessing the training set can improve the prediction efficiency of the model. FIG. 3 illustrates a data preprocessing method according to some embodiments of the present disclosure.

As shown in Figure 3, data preprocessing includes data preparation, feature processing, and feature engineering.

Feature processing includes missing value processing, feature screening, and data partitioning. The method of feature processing according to some embodiments of the present disclosure is introduced below.

In some embodiments, the risk prediction model further includes at least one of the following: filling the missing values in the user features with default values; filling the missing values in the user features with the mean value filling method; filling the missing values in the user features with the nearest completion method value;

Fill missing values in user features with nearest neighbor; fill missing values in user features with multiple imputation.

For example, for the samples in the training set and validation set, if the value of some features is missing among the multiple features of the sample, the missing feature value will be filled with the default value, and the missing value in the label will be retained without modification. In addition, in addition to filling missing values with default values, you can also use special value filling, average filling, hot card (nearest filling) method, nearest neighbor method, multiple imputation method, model prediction and other methods to deal with missing values.

In some embodiments, the training method of the risk prediction model further includes filtering out at least one of the following user features: user features whose missing rate is greater than a first threshold; user features whose discrimination degree to training samples is lower than a second threshold; variance greater than User characteristics for the third threshold.

For example, feature screening processing is performed on multiple features of the training samples and/or verification samples. Features with high missing rates and/or unstable features with high variance are excluded. It is also possible to exclude invalid features, that is, user features whose discrimination against training samples is lower than the second threshold, for example, constant-unity features with constant values. Among them, the degree of discrimination is an index to evaluate the ability of a feature to distinguish users. By using a single feature as the input of the risk prediction model, use AUC (Area Under Curve, area under the curve), KS (Kolmogorov-Smirnov, Lorenz curve), IV (Information Value, information value) to calculate the discrimination of features.

In addition, adopting feature dimension reduction methods, such as PCA (principal component analysis, principal component analysis), etc., can also realize the screening of effective features.

When dividing the training samples, the training samples are randomly divided into a training set (train) and a test set (test) according to a certain proportion, and the verification samples are not divided. Among them, the training set is used to train the model, and the test set is used to supervise the model. The training process prevents overfitting, and the validation set is used to evaluate the performance of the model.

In some embodiments, the method for training the risk prediction model further includes: discretizing user features. For example, binning and discretization are performed on features (such as continuous features and multi-category features).

By discretizing the features, it can promote the rapid iteration of the model, reduce the risk of model over-fitting, and improve the stability of the model.

In addition, in feature engineering, one-hot encoding is performed on the category features in the training set to obtain embedding vectors; data standardization (for example, data normalization) and logarithmic (log) transformation are performed on continuous features. , to get a dense vector. Similar processing is also done for the verification samples.

Among the user features, there may be irrelevant features, and the features may also depend on each other. The present disclosure can process data more flexibly through feature processing.

In step S3, the risk prediction model is trained according to the risk prediction results of multiple training samples and the corresponding risk labels.

For example, for each training sample, the loss function of the training sample can be calculated according to the risk prediction result of the training sample and the label of the sample. According to the loss function of all training samples, the model parameters corresponding to the minimum value of the loss function are solved to update the model parameters and train the risk prediction model.

In some embodiments, the risk prediction model is a multi-task prediction model such as Multi-TaskDNN (Multi-Task Deep Neural Networks, multi-task deep neural network model) or tree model.

If the risk prediction model adopts a tree-structured model, use the principle of an additive model and step-by-step calculation to iteratively build a new tree until the condition for the model to stop training is reached.

If a multi-task deep neural network model is used, the parameters of the iterative model are updated during the training process, and the model can be trained for several rounds (epoch), wherein each epoch is trained on the basis of the previous epoch. During the training of the neural network model, gradient descent + backpropagation will be adopted for training, and the parameters to be optimized will be updated. During optimization, the optimization method can be adjusted for specific task types, such as gradient descent, stochastic gradient descent, minibatch gradient descent, momentum (Momentum) stochastic gradient descent, Adagrad (adaptive gradient, adaptive gradient) method, Adam (adaptive gradient) method, moment estimation, adaptive moment estimation) method, etc.

The loss function adopts cross-entropy loss function, 0-1 loss function, Hinge (hinge) loss function, exponential loss function, logarithmic loss function, etc.

After the model training is completed, the model is evaluated by evaluation indicators such as Auc, recall rate (Recall), balanced F score (F1), and KS.

Fig. 4 shows a flowchart of training a risk prediction model using risk prediction results according to some embodiments of the present disclosure.

As shown in Fig. 4, training the risk prediction model using the risk prediction results includes steps S31-S32.

In step S31, for each training sample in at least one training sample, calculate at least one training The loss function for each training example in the sample.

In step S32, the risk prediction model is trained according to the loss functions of a plurality of training samples.

For example, if the training sample A includes both the first risk label Y ₁ and the second risk label Y ₂ , then the features of the training sample A are input into the model, and the prediction result p ₁ for whether the first risk occurs and the prediction result for the first risk The prediction result p ₂ of whether the risk occurs. Calculate the loss function of the training sample A according to Y ₁ , Y ₂ , p ₁ , and p ₂ .

If the training sample B includes only one risk label (for example, the first risk label Y ₁ ), and the features of the training sample A are input into the model, it is only necessary to calculate the loss function of training the training sample A according to Y ₁ and p ₁ .

In some embodiments, for each training sample in at least one training sample, according to the first risk label, the second risk label, the prediction result of the first risk prediction task and the prediction result of the second risk prediction task, at least one The loss function of each training sample in the training samples, including: for each training sample in the at least one training sample, according to the first risk label and the prediction result of the first risk prediction task, calculating each of the at least one training sample The first loss function of the training sample; for each training sample in the at least one training sample, according to the prediction result of the second risk label and the second risk prediction task, calculate the second loss of each training sample in the at least one training sample function; calculating a loss function of each of the at least one training sample according to the first loss function and the second loss function of each of the at least one training sample.

For example, according to Y ₁ and p ₁ , calculate the first loss function L ₁ of the training sample A, according to Y ₂ and p ₂ , calculate the second loss function L ₂ of the training sample A, according to the first loss function and the second loss function , to calculate the final loss function of the training sample A.

In some embodiments, calculating the loss function of each of the at least one training sample according to the first loss function and the second loss function of each of the at least one training sample includes: according to at least one training sample The sum of the first loss function and the second loss function of each training sample in , and calculate the loss function of each training sample in at least one training sample.

For example, the final loss function of training sample A is L ₁ +L2.

A risk prediction model can have more than two risk prediction tasks, and correspondingly, a training sample can also have more than two risk labels. Taking the multi-task neural network model as an example, the following describes how to train the model when there are N risk prediction tasks.

FIG. 5 illustrates a method of calculating a loss function according to some embodiments of the present disclosure.

For N risk prediction tasks, the corresponding risk labels are Y={Y ₁ , Y ₂ . . . Y _N }. Due to the hysteresis of labels relative to user features, some labels may be missing, so not every training sample can have all risk labels, that is, any risk label of a training sample can be a null value.

The risk prediction result of the risk prediction model for N tasks is P={p ₁ ,p ₂ ...p _N }.

For any training sample, traverse all the risk labels of the training sample, if the current risk label Y _i is empty, then skip, otherwise, according to Y _i and p _i , calculate the loss function of the training sample on the i-th task Loss _i . Sum all the loss functions of the training sample to get the total loss function of the training sample. Wherein, the loss function Loss _i of the training sample on the i-th task may be a cross-entropy loss function.

In some embodiments, the risk prediction model is validated using validation samples, wherein the validation samples are generated at different times from the training samples. For example, specify time as an observation point in the database, and determine the time window, generate a training set and a validation set across time. Among them, the training set is used to train the model, and the validation set is used to evaluate the performance of the model.

The verification sample is an OOT (Out of Time, across time) verification sample, which can be used to simulate and evaluate the effect of the online application of the model. The label generation time of the OOT validation samples is different from that of the training samples, therefore, the OOT validation samples are generated at different times from the training samples. For example, a training sample is a sample generated based on this year's data, and a validation sample is a sample generated based on last year's data. The time windows corresponding to the risk labels of training samples and validation samples do not overlap. Using cross-time verification samples to verify the model can ensure the stability of the model and improve the generalization ability of the model.

When using test samples to test the model and/or using verification samples to verify the model, the method of calculating the loss function is similar to the training process, and will not be repeated here.

According to the training method of the risk prediction model of the present disclosure, by obtaining multiple training samples, using the risk prediction model, for each training sample, according to user characteristics, a risk prediction result is generated, and then according to the risk prediction results of the multiple training samples and the corresponding risk labels, train the risk prediction model, use the first risk label and the second risk label of some training samples, and train the risk prediction model on multiple tasks at the same time, and obtain the first risk prediction task and the second risk prediction task. forecast result.

The risk prediction model not only learns the relationship between long-term risk and user characteristics, but also reduces the impact of risk label delay feedback on long-term risk prediction; it can also use more fresh training samples to learn the relationship between short-term risk and user characteristics. At the same time, it increases the number of training samples available. In addition, there is a certain correlation between the first risk prediction task and the second risk prediction task, and they have their own noises. This method of simultaneously training a risk prediction model on multiple risk prediction tasks can achieve data enhancement. Therefore, the trained risk prediction model can have higher accuracy in long-term risk prediction and short-term risk prediction.

In addition, the risk prediction model trained according to the risk prediction model training method of the present disclosure can simultaneously predict multiple risk prediction tasks, and has better generalization.

The present disclosure provides a risk prediction method, including: according to the user characteristics of the target user, using a risk prediction model to generate a risk prediction result for the target user, wherein the risk prediction model is based on the risk prediction model described in any embodiment of the present disclosure obtained by the training method. In some embodiments, the risk prediction method is performed by a risk prediction device.

Among them, the target user can be any user whose default risk needs to be assessed. After the risk prediction model is trained, according to the user characteristics of the target user, the risk prediction model can be used to generate the prediction result of the first risk prediction task and/or the prediction result of the second risk prediction task corresponding to the second risk label , so as to determine whether the target user will default in the first time window in the future and/or whether the target user will default in the second time window. Wherein, the user feature is extracted at a specified time, the first time window is a time window from the specified time to the first time, and the second time window is a time window from the specified time to the second time.

Fig. 6 shows a block diagram of a training device for a risk prediction model according to some embodiments of the present disclosure.

As shown in FIG. 6 , the risk prediction model training device 6 includes an acquisition module 61 , a generation module 62 and a training module 63 .

The acquisition module 61 is configured to acquire a plurality of training samples, wherein each training sample includes a risk label and user characteristics, and the risk label of at least one training sample in the plurality of training samples includes a first risk label generated in a first time window With the second risk label generated in the second time window, the length of the first time window is greater than the length of the second time window, for example, execute step S1 as shown in FIG. 6 .

The generation module 62 is configured to use a risk prediction model to generate a risk prediction result for each training sample according to user characteristics, wherein the risk prediction result of at least one training sample includes the prediction of the first risk prediction task corresponding to the first risk label The result and the prediction result of the second risk prediction task corresponding to the second risk label, for example, execute step S2 as shown in FIG. 6 .

The training module 63 is configured to train a risk prediction model according to the risk prediction results of a plurality of training samples and corresponding risk labels, for example, execute step S3 as shown in FIG. 6 .

The disclosure provides a risk prediction device, including a generation module. Wherein, the generation module is configured to use a risk prediction model to generate a risk prediction result for the target user according to the user characteristics of the target user, wherein the risk prediction model is obtained by training the risk prediction model training device described in any embodiment of the present disclosure .

FIG. 7 shows a block diagram of an electronic device according to other embodiments of the present disclosure.

As shown in FIG. 7 , the electronic device 7 includes a memory 71 ; and a processor 72 coupled to the memory 71 , and the memory 71 is used for storing and executing a training method of the risk prediction model. The processor 72 is configured to execute the method for training the risk prediction model in some embodiments of the present disclosure based on the instructions stored in the memory 71 .

According to some embodiments of the present disclosure, the risk prediction model training device and/or electronic equipment use the first risk label and the second risk label of some training samples to train the risk prediction model on multiple tasks at the same time, and obtain the first risk The prediction results on the prediction task and the second risk prediction task not only learn the relationship between long-term risk and user characteristics, reduce the impact of risk label delay feedback phenomenon on long-term risk prediction, but also use more fresh training samples. While learning the relationship between short-term risk and user characteristics, the number of available training samples is increased to achieve data enhancement. Therefore, the trained risk prediction model can have higher accuracy in long-term risk prediction and short-term risk prediction.

Figure 8 shows a block diagram of a computer system for implementing some embodiments of the present disclosure.

As shown in FIG. 8, computer system 80 may take the form of a general-purpose computing device. Computer system 80 includes memory 810, processor 820, and bus 800 that connects the various system components.

The memory 810 may include, for example, a system memory, a non-volatile storage medium, and the like. The system memory stores, for example, an operating system, an application program, a boot loader (Boot Loader) and other programs. System memory may include volatile storage media such as random access memory (RAM) and/or cache memory. The non-volatile storage medium, for example, stores instructions for executing the method for training the risk prediction model in some embodiments of the present disclosure. Non-volatile storage media include, but are not limited to, magnetic disk storage, optical storage, flash memory, and the like.

The processor 820 can be realized by means of discrete hardware components such as general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gates, or transistors. accomplish. Correspondingly, each module, such as the judging module and the determining module, can be implemented by executing instructions in the memory of the central processing unit (CPU) to execute corresponding steps, or can also be implemented by a dedicated circuit that executes corresponding steps.

Bus 800 may use any of a variety of bus structures. For example, bus structures include, but are not limited to, Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MCA) buses, Peripheral Component Interconnect (PCI) buses.

The computer system 80 may also include an input and output interface 830, a network interface 840, a storage interface 850, and the like. These interfaces 830 , 840 , and 850 , as well as the memory 810 and the processor 820 may be connected through a bus 800 . The input and output interface 830 can provide a connection interface for input and output devices such as a monitor, a mouse, and a keyboard. The network interface 840 provides a connection interface for various networked devices. The storage interface 850 provides connection interfaces for external storage devices such as floppy disks, U disks, and SD cards.

Aspects of the present disclosure are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus and computer program products according to embodiments of the disclosure. It should be understood that each block of the flowchart and/or block diagrams, and combinations of blocks, can be implemented by computer readable program instructions.

These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable device to produce a machine such that execution of the instructions by the processor produces the processes implemented in one or more blocks of the flow diagrams and/or block diagrams. device for the specified function.

The computer-readable program instructions, which may also be readable and stored in the computer-readable memory, cause the computer to operate in a specific manner to produce an article of manufacture, including implementing the process specified in one or more blocks of the flowchart and/or block diagram. function instructions.

The disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects.

Through the risk prediction model training method, risk prediction method and device, electronic equipment, and computer-readable storage medium in the above embodiments, the accuracy of the risk prediction model in long-term risk prediction and short-term risk prediction is improved.

So far, the risk prediction model training method, device, and computer-readable storage medium according to the present disclosure have been described in detail. Certain details known in the art have not been described in order to avoid obscuring the concept of the present disclosure. Based on the above description, those skilled in the art can fully understand how to implement the technical solutions disclosed herein.

Claims (15)

1. A method of training a risk prediction model, comprising:

obtaining a plurality of training samples, wherein each training sample comprises a risk label and a user characteristic, the risk label of at least one training sample in the plurality of training samples comprises a first risk label generated in a first time window and a second risk label generated in a second time window, and the length of the first time window is greater than that of the second time window;

generating a risk prediction result by using a risk prediction model according to user characteristics for each training sample, wherein the risk prediction result of at least one training sample comprises a prediction result of a first risk prediction task corresponding to a first risk label and a prediction result of a second risk prediction task corresponding to a second risk label;

and training a risk prediction model according to the risk prediction results of the training samples and the corresponding risk labels.

2. The training method of the risk prediction model according to claim 1, wherein training the risk prediction model according to the risk prediction results and the corresponding risk labels of the plurality of training samples comprises:

for each training sample in the at least one training sample, calculating a loss function of each training sample in the at least one training sample according to the first risk label, the second risk label, the prediction result of the first risk prediction task and the prediction result of the second risk prediction task;

and training a risk prediction model according to the loss functions of the plurality of training samples.

3. The training method of the risk prediction model of claim 2, wherein for each of the at least one training sample, calculating a loss function for each of the at least one training sample based on the first risk label, the second risk label, the predicted outcome of the first risk prediction task, and the predicted outcome of the second risk prediction task comprises:

for each training sample in the at least one training sample, calculating a first loss function of each training sample in the at least one training sample according to the first risk label and the prediction result of the first risk prediction task;

for each training sample in the at least one training sample, calculating a second loss function of each training sample in the at least one training sample according to a second risk label and a prediction result of a second risk prediction task;

calculating a loss function for each of the at least one training samples from the first and second loss functions for each of the at least one training samples.

4. The method of training a risk prediction model according to claim 3, wherein calculating a loss function for each of the at least one training sample from the first and second loss functions for each of the at least one training sample comprises:

calculating a loss function for each of the at least one training samples according to a sum of the first loss function and the second loss function for each of the at least one training samples.

5. The method of training a risk prediction model according to any one of claims 1 to 4, wherein the user features of the at least one training sample are extracted at a specified time, the first time window being a time window from the specified time to a first time, and the second time window being a time window from the specified time to a second time.

6. The method of training a risk prediction model according to any one of claims 1 to 4, further comprising at least one of:

populating missing values in the user feature with default values;

filling missing values in the user characteristics by using a mean filling method;

filling missing values in the user characteristics by using a near completion method;

filling missing values in the user characteristics by using a nearest neighbor method;

and filling missing values in the user characteristics by using a multi-interpolation method.

7. The method of training a risk prediction model according to any one of claims 1 to 4, further comprising screening out at least one of the following user characteristics:

user features for which the loss rate is greater than a first threshold;

the discrimination of the training samples is lower than the user characteristics of a second threshold value;

user features having a variance greater than a third threshold.

8. The method of training a risk prediction model according to any one of claims 1 to 4, further comprising:

and carrying out discretization processing on the user characteristics.

9. The method of training a risk prediction model according to any one of claims 1 to 4, further comprising:

the risk prediction model is validated using a validation sample, wherein the validation sample is generated at a different time than the training sample.

10. The training method of the risk prediction model according to any one of claims 1 to 4, wherein the risk prediction model is a multitasking deep neural network model or a tree model.

11. A method of risk prediction, comprising:

generating a risk prediction result for the target user by using a risk prediction model according to the user characteristics of the target user, wherein the risk prediction model is obtained by training according to the training method of the risk prediction model according to any one of claims 1-10.

12. A training apparatus for a risk prediction model, comprising:

the training system comprises an acquisition module, a processing module and a processing module, wherein the acquisition module acquires a plurality of training samples, each training sample comprises a risk label and a user characteristic, the risk label of at least one training sample in the plurality of training samples comprises a first risk label generated in a first time window and a second risk label generated in a second time window, and the length of the first time window is greater than that of the second time window;

the generating module is used for generating a risk prediction result by using a risk prediction model according to the user characteristics for each training sample, wherein the risk prediction result of at least one training sample comprises a prediction result of a first risk prediction task corresponding to a first risk label and a prediction result of a second risk prediction task corresponding to a second risk label;

and the training module is used for training a risk prediction model according to the risk prediction results of the training samples and the corresponding risk labels.

13. A risk prediction device comprising:

a generating module configured to generate a risk prediction result for the target user by using a risk prediction model according to the user characteristics of the target user, wherein the risk prediction model is obtained by training with the training device of the risk prediction model according to claim 12.

14. An electronic device, comprising:

a memory; and

a processor coupled to the memory, the processor configured to perform a method of training a risk prediction model according to any one of claims 1 to 10, or a method of risk prediction according to claim 1, based on instructions stored in the memory.

15. A computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement a method of training a risk prediction model according to any one of claims 1 to 10, or a method of risk prediction according to claim 1.

Images